It's not about a fear of losing the original, but losing each step in the process (which may be thousands of steps), having flexibility, and maintaining pixel information/quality throughout multiple edits.
Another way to think about it (in terms of programming) would be editing some code in an existing project, compiling it, and then losing the source code for your edits. Sure, you could keep patching the compiled version (not very easy) or use some kind of instrumentation to wrap/modify existing behavior, but if you can't easily modify the source of your past work (or just a single revision along the way), then all you're doing is piling more patches (in compiled form) on top of old work - not very efficient, and you lose flexibility.
In a non-destructive workflow, you may (for example) adjust the color balance of a single object in the scene, keep working, and then realize that the color adjustments are out of place. Since that adjustment is preserved and editable at any time independently of all other edits, then it only requires an adjustment of the parameters - and you don't lose any information.
If you were to do that kind of edit directly to the pixels (even if they're on a separate layer for the object in question), then each time you adjust color/levels/etc, you're losing information that was originally there. Similar to resizing an image very small and then trying to enlarge it to original size, destructive color edits will start producing nasty artifacts like color banding (you'll start seeing a "combing" effect in the histogram when adjusting color even after a single previous color adjustment).
If you're worried about messing up a file, why not just make a copy of it before you edit it?